China leans on homegrown AI models to sharpen extreme-weather forecasts

China has found a new route in its push to become a global AI powerhouse as the country’s weather agency and research institutes have put domestically built AI models at the frontline of its typhoon monitoring and forecasting system. 

Asia’s largest economy also already has its eyes on expansion routes as it is now planning to spread the AI systems already in use in the country to other developing countries, especially those at extreme weather risk in the Asia-Pacific.

China is building an AI weather system for the world 

While China itself was the first customer of its domestic AI weather labs, the country has also signaled an ambition to allow them to spread across the world, potentially becoming the global standard. 

According to a Xinhua report carried by the State Council Information Office, the China Meteorological Administration (CMA) introduced the Fenghe weather services large language model at this year’s World Artificial Intelligence Conference in Shanghai as a global open-source project. 

The Fenghe model pushes out personalized weather information and risk alerts for Chinese users. 

The United Nations’ Early Warnings for All initiative received the international version, which offers consultations in Chinese and English and allows developers to build their own applications on top of the model.

Which countries are already using Chinese AI weather models?

Pakistan, Djibouti, Ethiopia, Jordan, Mongolia, the Solomon Islands, and Sri Lanka are seven of the developing countries that have already integrated Chinese AI weather models directly into daily disaster operations. 

However, they are running a model different from Fenghe. Instead, they are running Mazu (Multi-hazard Alert, Zero-gap and Universal), another system from the CMA, which is further along in its popularization agenda.

Pan Jinjun, a CMA chief engineer, called Mazu an “international public good” that delivers “tailored early warning services for developing nations.” 

Another 40 Global South countries are already experimenting with a quicker public cloud version of the system named after the sea goddess.

Can AI weather systems improve typhoon monitoring? 

The research side is producing measurable gains. Scientists at the Institute of Atmospheric Physics under the Chinese Academy of Sciences, working with Fudan University, paired China’s FuXi AI weather model with a method they call Orthogonal Conditional Nonlinear Optimal Perturbations, or O-CNOPs, to map the range of paths a storm might take. PreventionWeb reported the work on August 1.

Across 62 typhoon cases and 91 comparative experiments, the combined FuXi-CNOPs system matched the best global operational ensembles at 24 hours and beat them from 24 to 120 hours out, trimming maximum track errors by as much as 32.33% and tightening uncertainty estimates by up to 29.2%, according to the team led by research fellows Duan Wansuo and Li Hao. 

The study appeared in Advances in Atmospheric Sciences. Duan told the Global Times the approach needs no extra training of large models, which holds down computing costs and makes it easier to deploy in day-to-day operations.

A stack of Chinese models feeding the push

These tools sit on top of several Chinese forecasting models that have drawn international notice. FengWu, built by Shanghai researchers, produces global forecasts at 0.25-degree resolution. 

In a study published in Communications Earth & Environment, it outran the European Centre for Medium-Range Weather Forecasts’ high-resolution model along with Pangu-Weather and GraphCast, extended skillful forecasts past the 10-day mark, and cut five-day tropical cyclone track error to 201 kilometers for 2022.

Huawei’s Pangu-Weather, another homegrown system, generates forecasts far faster than traditional numerical prediction, and Huawei says both the ECMWF and the CMA’s National Meteorological Center have confirmed its advantages. Together, the models give Beijing a domestic base to draw on as it pitches its weather AI abroad.